Why some companies are abandoning use of AI coding assistance and why it is now the time to use coding assistance/assistants.

“An interesting trend I’m seeing is that some of the most elite teams are abandoning AI coding altogether.”

I hear this is happening in some companies and figure it is because:
(a) They either started too soon (before LLMs were smart enough;
(b) They picked the wrong coding assistant framework – usually because early on, coding assistant fork/extension that use LLMs were not advanced enough, [cough] Github Copilot);
(c) They did not use experienced-enough engineers to manage/use the AIs properly (establishing project framework, chunking the work, strict instructions/guidance for the underlying LLMs to abide by principles like modularity, naming conventions, etc.).

My Journey

Here’s a summary of my journey and why I think using AI now to code-assist and even write most or all of the code in a small or large project is viable, for now *requiring* it be done with the guidance of a senior programmer who has had the right kind of experience.

Qualification: 30 years freelance coding and before that, a few years as a kid.

3.5 years ago when I started playing with a beta of Github Copilot extension for VS Code, I deplored the code it wrote and how unaware it was of the surrounding code-base and overall app structure/intent. “Ick” and sadness. Uninstall.

Then models improved and competing coding assistant VS Code extensions like CLine, et al, came out. I played with many and *used* CLine, finding that out-of-the-box, it probably x3’d my productivity after accounting for many reverts.

Then I discovered a fork of CLine called Roo Code! Jumped immediately into it and wow! I don’t know if it was anything different about Roo (from CLine) or if it was because models were also improving. My productivity jumped to around x5. It helped that Roo has a thriving Discord community with devs that are hyper-attentive.

THEN I started writing custom instructions for built-in Roo modes/agents (think Architect, Orchestrator, Coder, etc.) and started adding my own modes/agents, like Coder Jr, Planner Team (more advanced architect that uses 3 chained modes), Debugger, Tester, Docs Writer, Githubber, etc.

And since then:
Turned it into a free repo called AgentFlow that I use in everyday coding of small to large projects and yeah, I’m guessing my productivity has skyrocketed to at least x10, maybe more! The FRAMEWORK creates reliability and accuracy!